I added a macro agent to the paper-trading desk. One of its jobs is to save the Cleveland Fed's inflation estimate exactly as it appeared that day. When the official CPI number comes out, regular software saves that too and checks how far off the forecast was. Later revisions stay separate so we do not grade an old prediction using cleaned-up history.
The most important setting is pretty simple:
trade_authority=false. It can talk. It cannot trade.
Guessing inflation correctly does not mean it guessed Bitcoin correctly. Those are two different bets, so I score them separately.
The macro agent can add context. It cannot declare the whole market bullish, decide a position size, overrule another strategy, or place an order.
Seeing something is not permission to act
The dashboard shows the forecast and how many old forecasts have a real answer now. Until there are enough of those, the honest label is not enough evidence. Boring, yes. Better than pretending the agent has a superpower.
I save three different things:
- What it saw. The source, time, target month, and a fingerprint proving the page did not change later.
- What it predicted. A number we can grade later, not a fuzzy paragraph.
- What it is allowed to influence. A software rule decides that. The agent cannot promote itself.
Do not grade it with tomorrow's answer
Government numbers get revised. If I compare an old forecast with a nicer number published later, I am giving the system information it did not have at the time. So the first official release is the one that counts. Revisions are saved, but they do not rewrite the score.
The same rule applies everywhere: if a bot says it predicted something, save what the world looked like before the answer was known.
The rule I am using
Save the input. Save the prediction. Grade it later. Keep “not enough evidence” visible. And never let a smooth explanation turn itself into permission.
This is all simulated trading. It is not financial advice, and the agent cannot reach a real brokerage account.